|
Abstract:
|
Spatially misaligned data are typically combined by choosing a common reference scale in order to conduct regression modeling, but this can cause significant measurement error in predictions. To overcome this issue, we developed a spatial functional regression approach to predict a response surface based on combining data referenced on different spatial supports, but related to each other by fixed distances. This model predicts a functional response based on functional predictors, where the response and predictors are bivariate, non-parametric functional surfaces belonging to a 2D functional space. We apply this method to predict a surface of PM2.5 air pollution concentrations from Multi-angle Imaging SpectroRadiometer (MISR) aerosol optical depth (AOD) satellite observations, and meteorological parameters. Here, the outcome (PM2.5) is measured at point locations, while the predictors are observed on a grid (AOD), and at a different set of point locations (meteorology). By cross validation, our results show that predicted PM2.5 concentrations based on the functional regression approach have less uncertainty than those generated from geostatistical techniques such as kriging.
|
